Post on 15-Aug-2020
transcript
Institut Mines-Télécom
Satellite Image Mining :
Indexing and Retrieval
Master AIC (Apprentissage, Information et Contenu) and D&K (Data & Knowledge) – Université Paris Saclay
Henri Maître - Télécom ParisTech
henri.maitre@telecom-paristech.fr
Institut Mines-Télécom
Content
I - Remote Sensing (RS) and RS Images ……………………………..
• Why remote sensing? ……………………………………………….
• Preparing a RS program …………………………………………..
• Image parameters (resolution, spectral bands, repetition …)• Image diversity (12 not presented slides) ……………………….
II- RS Image mining ………………………………………………………
• RS archiving problems …………………………………………….
─ RS image mining IS NOT multimedia image mining …………
─ RS image mining specificity ……………………………………….
• Hand-crafted features and classification ………………………..
─ Expert in the loop, relevance feedback ………………………..
• Deep Neural Networks ……………………………………………...
─ DNN toolbox ……………………………………………………….
─ Some instances ……………………………………………………
• From Lo to Hi level of semantics ………………………………….
3
4
8
14
23
25
26
28
35
45
57
60
64
67
72
09/01/20192 Master AIC
Institut Mines-Télécom
Part I - Remote Sensing and
Remote sensing images
Why? How? For Whom?
09/01/20193 Master AIC
Institut Mines-Télécom
Why do we need Remote Sensing
Environnement:
• Meteorology: short-term weather prediction
• Climate: long-term monitoring
• GMES = Global Monitoring for Environment and Security: survey of
natural and man-made catastrophies
─ volcanos
─ earthquake, tsunamis, floods
─ Industrial hazards
─ Marine pollution
4 09/01/2019 Master AIC
Institut Mines-Télécom
Why do we need Remote Sensing
Agriculture :
• Survey and evaluation of crop & farming production
• Fish & Aquaculture resources management
• Forestry resources planning
• Water management, dams, watering
• Desertification & urban pressure
5 09/01/2019 Master AIC
Institut Mines-Télécom
Why do we need Remote Sensing
Town & country planning:
• Mapping and inventories
• Constructions & public work: railways, airports, harbours, dams, …
• Cities and Mega-cities management
• Management of moving populations, displacements, installation
• Climatic impact management
• Crisis management: fires, floods, …
6 09/01/2019 Master AIC
Institut Mines-Télécom
Why do we need Remote Sensing
Defence & Security applications:
• Military deployment preparation
• Military mission debriefing
• Intelligence and survey of national/foreign territory
7 09/01/2019 Master AIC
Institut Mines-Télécom
How to prepare a remote sensing program
09/01/20198 Master AIC
Institut Mines-Télécom
How is prepared a remote sensing program
Conceive the sensor: application, customers
Determine which satellite / which launcher
Conceive the ground-station and the data management
process : economical, social and technical issues
15 to 20 years
9 09/01/2019 Master AIC
Institut Mines-Télécom
Terrestrial linkAerial link
Telemetry link1.6 Mbps
S band Station Emiter/Receiver
Telecontrol link4 kbps
S Band
User
X Band
Image down link250 Mbps
X band receiving Station
Processing center
Images
ImagesRequest for an image
Acquisition
ProgrammingOperating and
Control Decision Center
S-Band Antenna
X Band Antenna
Satellite links with the Earth
09/01/201910 Master AIC
Institut Mines-Télécom
Satellite : orbit choice
Mecanics laws:
• Newton = centripetal force
• Satellite speed = driving force
elliptical or circular trajectory (Kepler)
r
r
r
mF
2
11 09/01/2019
r
Earth
12 742 km
Processing
satellite : 800 kmGeostationary
36 700 km
Master AIC
Institut Mines-Télécom
Orbit choice
1) Geostationary
• Always in the Equator Plane
• Always at vertical of the same point on the Equator
• Altitude ~ 36 700 km
• Field of view: ~1/3 Earth: always the same
• Applications : meteo, survey of catastrophies, telecoms, TV
12 09/01/2019 Master AIC
Institut Mines-Télécom
Orbit choice
2) Processing satellite (low orbit)
• Altitude ~ 800 km (down to 250 km)
• Circular ~ N/S
• Trajectory : ± polar
• ~ 15 revolutions / day
• Helio-synchronous
13 09/01/2019 Master AIC
Institut Mines-Télécom
Choice of resolution
Pixel size = smallest measured terrain
on the ground
• from 30 cm to 10 km
14 09/01/2019
Δx = 2,5m
SPOT 5
Master AIC
Institut Mines-Télécom
On Ground resolution
Depends on:
• Sensor :
Photosites size: 𝛿𝑥
𝐺 =𝑓
𝐷= enlargement
∆𝑥 = 𝛿𝑥
𝐺= smallest detail
• The camera lens
δ′𝑥 =λ𝑓
𝑑= diffraction limited resolution
∆𝑥𝑚𝑖𝑛 =λ𝑓
𝐺𝑑= λ𝐷
𝑑
D = satellite-Earth distance
~ 1 000 km = 106 m
λ = wave length
= 0,5 . 10-6 m
d = lens diameter
~ 0,5 m
∆𝒙𝒎𝒊𝒏 = 𝟏𝒎
Possible with : f = 1 m
if δ′𝑥 =𝝀𝒇
𝒅= 1µm
the photosite measures 10-6 m
Smallest detail
15 09/01/2019 Master AIC
Institut Mines-Télécom
Often push-broom sensor
Sensor size along track:
• On line sensor
• = speed x aperture time
In the other direction
• Number of sensors on a linene
• from 6 000 to 40 000
Resolution :
• Depends on the lens
16 09/01/2019 Master AIC
Institut Mines-Télécom
Swath choice
Swath = image width
• from 10 km to 10 000 km
• = from 3 000 to 40 000 pixels / line
• Given by the sensor size
• Limited by the communication link with Earth
Revisit delay
15 min for geostationnary sat. (to dump the memory)
• from 1h30 (min) to 1 month for processing satellites
• But … sensor agility!
17 09/01/2019 Master AIC
Institut Mines-Télécom
Video possibility
Angle of view ~ + or – 50 degrees:
• MN ~ 2000 km
• 1 rotation around the Earth = 90 min
~ 40 000 km
• Time to go from M to N
= 90*2000/40000 = 4 min 30 s
18 09/01/2019
M
N
Master AIC
Institut Mines-Télécom
Which wave length?
1 – Passive sensors: measure the energy sent back from Sun
by Earth or the energy radiated by Earth
• Emitted from the Sun (Wien’s law) x Atmosphere transparency x
Ground Reflexion
• Black and White (Panchromatic)
• Visible = Blue - Green - Red
• Visible and Near Infra-Red : G - R - IR = false colors
• Multispectral : 7 20 channels
• Hyperspectral : 64 512 channels
© Wikipedia © Wikipedia
19 09/01/2019 Master AIC
Institut Mines-Télécom
False colors : NIR-R-G R-G-B
20
vegetation = red
09/01/2019
False colors True colors
Master AIC
Institut Mines-Télécom
Multispectral image visualisation:
pseudo colors
09/01/201921
Landsat = 7
channels
© UVED
321 432
754
542 435
1
2 R
3
4 G
5
6 + B
7
41(7+5)
Master AIC
Institut Mines-Télécom
Which wave length?
2 – Active sensors: EM emitter + receiver
radar = Micro waves: λ= 1 cm to 10 m
• But low resolution ∶ ∆𝑥 =λ𝑓
𝐺𝑑
• With complex processing: Synthetic Aperture Radar hi resolution
22 09/01/2019 Master AIC
Institut Mines-Télécom
One point is seen from several antenna positions
When moving, superposition of the seen areas
Real antenna is too small, it covers a very large field
From computation we obtain an accurate information = synthetic
antenna09/01/201923 Master AIC
Institut Mines-Télécom
Satellite images = big data !
Television HD 1 280 x 720 pixels
Television 4k 4 000 x 2 000 pixels
PC display screen 1 600 x 1 200 pixels
Photo camera 5 000 x 4 000 pixels
Spot 1 … 4 6 000 x 6 000 pixels
SPOT 5 24 000 x 24 000 pixels
Quickbird 40 000 x 40 000 pixels
1 600 000 000 pixels = 1,6 Gpixels
= 800 PC display screens
1 SPOT 5 image = 10 s of satellite run
24 09/01/2019 Master AIC
Institut Mines-Télécom
Diversity of Remote Sensing Images
(slides are not presented in the lecture
notes)
09/01/201925 Master AIC
Institut Mines-Télécom
Part II – Remote Sensing
Image Mining
09/01/201926 Master AIC
Institut Mines-Télécom
Remote Sensing Imaging: Archiving Problems and Issues
09/01/201927 Master AIC
Institut Mines-Télécom09/01/2019
Satellite Image archives
How can we store millions of images?
How can we ensure durability of storage?
How knowing that information exists?
How retrieving information?
How exploiting information?
Data Mining directly on image
files
When searching in a small set of
images
Indexing images when received
data mining on index
When searching in large sets
28 Master AIC
Institut Mines-Télécom
RS Image mining IS NOT MultiMedia Image Mining
09/01/201929 Master AIC
Institut Mines-Télécom
Mining in Multimedia Image databases
Multimedia information retrieval :• Either from semantic information: name, description, caption, text
(90 % of Google-like retrieval)
• Or from instance (i.e. with a reference image)
(Face or fingerprint recognition)
I – Classical Machine Learning techniques (2000-2012)• Hand-crafted feature detection and/or salient point detection
• Classification in p-dimensional space─ few parameters
─ few learning images (groundtruth) ~ 1000
II – Deep neural networks (2012 - …) • Directly with images as input and/or with extracted features
• Several +/- linear classifiers in cascade ─ thousands of parameters
─ hundred of thousands of images as groundtruth
09/01/201930 Master AIC
Institut Mines-Télécom
Multimedia image mining: handcrafted
features + classification
Multimedia information retrieval from exemple:
• Choices: to be robust to possible differences
─ scale, lighting, orientation, color, … invariance
• Strategy: detect invariant features
─ Histograms, color distribution, area-based segmentation, graph
description, …
─ Textures
─ Salient point detection: Harris, SIFT, SURF, …
• Represent the image as a vector in a p dimensional space ℝp
• Classification : Bayès, k-NN, dynamic clustering, SVM (Support
Vector Machine), Graph tree, …
09/01/201931 Master AIC
Institut Mines-Télécom
Ambiguous semantics: Venus
09/01/201932 Master AIC
Institut Mines-Télécom
Textual categorisation
09/01/201933 Master AIC
Institut Mines-Télécom
invariance
09/01/201934 Master AIC
Institut Mines-Télécom
Salient points: SIFT
09/01/201935 Master AIC
Institut Mines-Télécom
Specificities of RS Image mining
09/01/201936 Master AIC
Institut Mines-Télécom 09/01/2019
Category-based retrieval in specific data-bases
Examples:
• Biomedical
• Biology
• Astronomy
• Remote sensing and satellite images
Goal: to retrieve images « looking the same » as a given sample in very specialized data-bases
Different from : retrieving the exact object in a very broad data-base
Institut Mines-Télécom09/01/2019
Satellite images
A very specific content
Fields
City
Forest
38 Master AIC
Institut Mines-Télécom
A same region, different signals
09/01/201939
From : Tong et al.
arXiv 1807.05713 - 2018
Master AIC
Institut Mines-Télécom
The role of scale
09/01/201940
High-Badakchan, Tadjikistan - Ikonos
15 m 1 m
Master AIC
Institut Mines-Télécom
Main scales
<1 meter = Very high resolution : fine details in urbancontext, cars, pedestrians, containers, fences, small boats, …
1 m < … < 5 m = High resolution : urban fine structures, houses, streets, gardens, individual trees, railway & road networks, …
5 m < … < 30 m = Middle resolution: fine landcover, coarseurban structure: dense urban, residential or commercial areas,
> 30 m = low resolution: global landcover
09/01/201941 Master AIC
Institut Mines-Télécom
Available information on satellite images
(semantic information) = Ancillary data
Accurate positionning in universal geographical
references: UTM, Mercator, Lambert, etc…
Precise time referencing: seasonal variations (vegetation,
insolation, agricultural production, …), sun positionning
(shadows), tide effects (precise coast-line, harbours and
fishering activities), meteorological conditions (snow, floods,
…)
Satellite parameters: resolution, spectral sensitivity, noise
Often: Image quality, cloud cover
09/01/201942 Master AIC
Institut Mines-Télécom
Satellite images
What are we looking for? It is not clear!
• Precise objects:
─ A boat a road-crossing
─ A building an airplane landing area
• Generic objects:
─ A marina a forest fire
─ Greenhouse cultures refugee camps
─ Oil pipeline typhoon hazards
─ A geological synclinal
• Specific terrain configurations:─ Conducive to: … floods, … desertification, … urban pollution, …
─ Conducive to: … build a factory, … plan a bombing, … cultivate marijuana
09/01/201943 Master AIC
Institut Mines-Télécom09/01/2019
Spatial scale vs. Semantic complexity
pixels regions zones
Spatial scale
Semantic Complexity
edges
roads
field
intensive
farming
house
village
middle-age
city
school
flower
culture
car
geographic
landmark
mixed field
agriculture
greenhouses marina
44 Master AIC
Institut Mines-Télécom
Hierarchical representation
spectral properties (R,G,B,IR)
Pixel contrast / texture
edges, contours
Objects Scene
form / shape
Region
content (spectral : textural)
09/01/201945
Increasing semantics
sea
wharf
warehouse
house
network
fields
Master AIC
Institut Mines-Télécom
RS image processing & hand-crafted feature detection
09/01/201946 Master AIC
Institut Mines-Télécom
Mining in RS Image databases
Semantic information retrieval :• From ancillary data
I – Classical Machine Learning techniques (2000-2012)• Image Processing
• Hand-crafted feature detection and/or salient point detection
• Classification in p-dimensional space─ few parameters
─ few learning images (groundtruth) ~ 1000
II – Deep neural networks (2012 - …) • Directly with images as input and/or with extracted features
• Several +/- linear classifiers in cascade ─ thousands of parameters
─ hundred of thousands of images as groundtruth
09/01/201947 Master AIC
Institut Mines-Télécom09/01/2019
Probabilistic evaluation
p(Li|w)
48 Master AIC
Institut Mines-Télécom09/01/2019
Hand crafted features
Radiometry• Multispectral : channels
• Specific combinations for remote sensing : NDVI + IB + ISU
Textures• Gabor Filters
• Haralick cooccurrence matrices and their descriptors
• Quadratic Mirror Filters (wavelets)
• Contourlet decomposition
• Steerable wavelets
• Markov random fields parameters (Gaussian, Laplacian, Log-laplacian …)
Structures• Contours, regions, lakes, forests, deserts
• Objects : roads, buildings, rivers, lakes
• Roads, Train or River networks
49 Master AIC
Institut Mines-Télécom09/01/2019
Some efficient choices
Indexing: small subimages: (~ 64 x 64 pixels) = 320 m x 320 m on the groundfor SPOT 5 images
Mixed features:
• Radiometry (Panchro only)
• Structure (contours)
• wavelets : 2 directions, 4 scales
Automatic feature selection (supervised: ReliefF, Fisher FS, SVM-RFE or
Unsupervised: MIC (Max Information Compression), k-means FS)
~ 100 features with or 10 to 20 features
redundancy without redundancy
Give names to classes (from label to name)
• Waste fields
• Cultures
• Housing
• Road and river networks
50 Master AIC
Institut Mines-Télécom09/01/2019
Classification
label = 24
or
name = « Corn field »
Semantic labelling
Many different classifiers:
• MAP & Bayes decision
• K-nearest neigbours
• Graph tree
• Kernel methods (SVM = Support Vector Machine)
• Hierarchical clustering
Supervisedor
Unsupervised
51 Master AIC
Institut Mines-Télécom 09/01/2019
Supervised classes
Residential
areas
Planes
Industrial
tanks & cisterns
Railway
marshalling yard
Institut Mines-Télécom09/01/2019
Supervised classes
factories
Dense urban
area
villages
Urban parks
53 Master AIC
Institut Mines-Télécom 09/01/2019
Supervised classes
Graveyards
Road
interchange
Castle
parks
Parking lots
Institut Mines-Télécom
How to express results?
Classification rate 97.3 % (or error rate: 2.7 %)
Confusion matrix
Receiver Operating Characteristic (ROC Curve)
Convert TP and FP into FPR and TPR ϵ [0,1]
Plot TPR = f(FPR) for many different parameters
Without specific instruction, take the closest
point from A = (0,1) as working condition
09/01/201955
Present object Absent object
Positive detection True positive (TP)False positive (FP)
(type I error)
Negative detectionFalse negative
(type II error)True Negative
TPR
FPR
OO
1
1
A
Master AIC
Institut Mines-Télécom09/01/2019
Sub image classification (128 x 128) :
city, wood, fields, sea, desert & clouds
600 images for each class
Results: Gaussian SVM,
Mean error 1.4% ± 0.4%
(147 features, cross validated)
True\Found
(%)
city clouds desert fields woods sea
city 98.8 0 0 0.5 0 0
cloud 0 99.3 0.2 0 0 0
desert 0 0 99.0 0.3 0 0
fields 0.5 0.2 0.8 98.1 0.3 0.4
woods 0 0.2 0 0 98.0 1.4
sea 0.7 0.3 0 1.0 1.7 98.2
Typical performances of algorithms
56 Master AIC
Institut Mines-Télécom09/01/2019
How many features?
Automatic feature selection
• Wrappers
• Filters (mutual information)
• Embedded (Lasso)
57 Master AIC
Institut Mines-Télécom
Using a human expert to improve learning
09/01/201958 Master AIC
Institut Mines-Télécom 09/01/2019
Learning with Relevance feedback
Man Machine dialog
Subjective Objective
Institut Mines-Télécom 09/01/2019
Database composed of 600 SPOT5 images divided in 6 classes
Used features: Gabor, Haralick, QMF and GMRF
Gaussian Kernel
System evaluation: Precision-Recall graphs
Institut Mines-Télécom
Deep Neural Networks
09/01/201961 Master AIC
Institut Mines-Télécom
Mining in RS Image databases
Semantic information retrieval :• From ancillary data
I – Classical Machine Learning techniques (2000-2012)• Image processing
• Hand-crafted feature detection and/or salient point detection
• Classification in p-dimensional space─ few parameters
─ few learning images (groundtruth) ~ 1000
II – Deep neural networks (2012 - …) • Directly with images as input and/or with extracted features
• Several +/- linear classifiers in cascade ─ thousands of parameters
─ hundred of thousands of images as groundtruth
09/01/201962 Master AIC
Institut Mines-Télécom
Some references (dated 01/01/2019)
Zhang, L., Zhang, L., & Du, B. (2016). Deep learning for remote sensing data: A technical tutorial on the state of the art. IEEE Geoscience and Remote Sensing Magazine, 4(2), 22-40.
Maggiori, E., Tarabalka, Y., Charpiat, G., & Alliez, P. (2017). Convolutional neural networks for large-scale remote-sensing image classification. IEEE Transactions on Geoscience and RemoteSensing, 55(2), 645-657.
Han, J., Zhang, D., Cheng, G., Guo, L., & Ren, J. (2015). Object detection in optical remote sensing images based on weakly supervised learning and high-level feature learning. IEEE Transactions on Geoscience and Remote Sensing, 53(6), 3325-3337.
Tong, X. Y., Lu, Q., Xia, G. S., & Zhang, L. (2018). Large-scale Land Cover Classification in GaoFen-2 Satellite Imagery. arXiv preprint arXiv:1806.00901.
Boualleg, Y., & Farah, M. (2018, July). Enhanced Interactive Remote Sensing Image Retrievalwith Scene Classification Convolutional Neural Networks Model. In IGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium (pp. 4748-4751). IEEE.
Marmanis, D., Datcu, M., Esch, T., & Stilla, U. (2016). Deep learning earth observation classification using ImageNet pretrained networks. IEEE Geoscience and Remote SensingLetters, 13(1), 105-109.
Kussul, N., Lavreniuk, M., Skakun, S., & Shelestov, A. (2017). Deep learning classification of land cover and crop types using remote sensing data. IEEE Geoscience and Remote SensingLetters, 14(5), 778-782.
Penatti, O. A., Nogueira, K., & dos Santos, J. A. (2015). Do deep features generalize fromeveryday objects to remote sensing and aerial scenes domains?. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops (pp. 44-51).
09/01/201963 Master AIC
Institut Mines-Télécom
Deep Neural Network
09/01/201964
From: I. Bloch, AIC
? ?
Which input?
Raw image
Processed image (filtered, segmented …)
Feature detected image (classified, edgedetected, …)
Features
Which output?
Densely classified image
Detected targets
List of targets
List of Feature
Which architecture?• # layers,
• type of layers
Which protocole?• Feature learning
• Fine tuning
Master AIC
Institut Mines-Télécom
CNN basic components
09/01/201965
Convolutional layer: with rxr kernel – down
scaling
Nonlinearity: sigmoïd or RELU (rectified linear
unit)
Pooling layer: single value taken from
a set of values - ex: max on a rxr patch
Autoencoder: symetrical NN to reduce the
model dimensionality
Master AIC
Institut Mines-Télécom
CNN basic components
09/01/201966
Fully convolutional layer: to perform a large
distance context dependance
Transfer coding: to learn from a database and
use for another one
Fine Tuning: to specify a network to a given
task after training on a general purpose data
base
Yoyo architecture : downsampling for feature
extraction then upsampling for fine positioning
of targets
Master AIC
Institut Mines-Télécom
Most used components for RS-CNN (2019)
CNN from the Pattern Recognition community
• AlexNet
• GoogleNet
• VGGNet
• ResNet
• Inception
Training sets
• ImageNet (General purpose image library for pattern recognition)
• UC Merced DataSet (Aerial images / 21 classes)
• OSM - OpenStreetMap (Aerial Image Database)
• Google Street Map (hi level semantic)
• NLCD - USGS data Base (Geological survey)
• Corinne Landcover (Agriculture & vegetation)
• Gaofen Image Dataset (GID) (Hi Resolution Satellite)
• …
09/01/201967 Master AIC
Institut Mines-Télécom
Instance # 1 : Basic CNN (DLR)
With UC Merced Land database (aerial / 21 classes)
With pre-trained CNN (Imagenet)
Fine-tuned full convolutional layers with enhanced data
09/01/201968
Marmanis et al. IEEE TGRS,
Jan 2016
Master AIC
Institut Mines-Télécom
Instance # 2 : fully CNN (Inria)
09/01/201969
Patch-based CNN Fully convolutional Patch -based CNN
Image ground truth patch based fully convolutional SVM
Maggiori et al. IEEE TGRS, feb 2017
Master AIC
Institut Mines-Télécom
Instance # 3 : RS CNN (Liemars/Wuhan)
09/01/201970
From : Tong et al.
arXiv 1807.05713 - 2018
Master AIC
Institut Mines-Télécom
Instance # 3 : RS CNN (Liemars/Wuhan)
09/01/201971
From : Tong et al.
arXiv 1807.05713 - 2018
Cooperation between classifying (sparse) and segmenting (dense)
Master AIC
Institut Mines-Télécom
Instance # 3 : RS CNN (Liemars/Wuhan)
09/01/201972
From : Tong et al.
arXiv 1807.05713 - 2018
Master AIC
Institut Mines-Télécom
From Low to High Level - Changing the scale
09/01/201973 Master AIC
Institut Mines-Télécom
Complexity of images
09/01/201974
Analysis window : real size
128 x 128 pixels
Analysis window : enlarged
Master AIC
Institut Mines-Télécom09/01/2019
Hierarchical representation
pixels regions zones
Spatial scale
Semantic Complexity
edges
roads
field
Intensive farming
house
village
Middle-age city
school
Flower culture
car
Geographic landmark
Mixed field agriculture
greenhousesMarina
Two goals:
• Enlarge the field of view
• Increase the semantic level
Grouping strategy:
• Sliding window
• Pyramid
• Growing and Merging
Decision strategy:
• Bag of Visual Words (BOVW)
75 Master AIC
Institut Mines-Télécom
Increasing the semantics
09/01/201976
Park = {trees+fields+tracks}
Waste area ={waste+lawns+trees+roads}
Residential area = {houses + lawns + pools + roads}
Commercial area = {buildings+houses+parking lots+ waste
Master AIC
Institut Mines-Télécom09/01/2019
Probabilistic evaluation
p(Li|w)
77 Master AIC
Institut Mines-Télécom
Decision making: Bag of Words
2 levels H=high (unknown) L = low (known)
List of N classes at H = {c1,c2,… cN}
At H : 1 super-region with n objects, each ∈ 1 class = n
labels described by the ordered list of the probability (or the
occurrence) of each class:
Rk={p1,p2, …pn}
Classify H according to the Rk
• Naïve Bayes : 𝒄∗= argmax 𝒑 𝒄 𝒙 = argmax 𝒑 𝒄 ς𝒌=𝟏
𝒏 𝒑 𝒙𝒌 𝒄
• Improving Naïve Bayes:
─ pLSA = Probabilistic Latent Semantic Analysis
─ LDA = Latent Dirichlet Analysis
09/01/201978 Master AIC